In the early days of AI-assisted document authoring, most teams started with prompt libraries — and that made sense. They’re easy to get started with, sound intuitive to end users, and can quickly automate parts of a workflow.

Need to automate writing a Clinical Study Report? You experiment with different prompts for generating each section and sub-section and soon end up with a master library of “CSR prompts.” But over time, these libraries grow and become cumbersome. You realize you need a different version of the prompt depending on the study phase, therapeutic area, study design, and more. You also need to tweak prompts depending on which LLM you’re using (Claude 3.7 vs. GPT-4o, for example).

Suddenly, scaling prompt libraries across the thousands of documents your organization produces becomes… hard.

At Artos, we chose a different path: an agentic, modular architecture that actually gets better as it scales.

The Limits of Prompt Libraries

Prompt libraries look appealing on the surface. They’re simple, fast to prototype, and easy to tweak. But in enterprise settings — especially in life sciences — they become brittle, fast:

• They expect too much from users. End users are suddenly expected to master prompt engineering on top of their core responsibilities as domain experts.

• They’re hard to manage. Keeping hundreds of prompts version-controlled, up to date with changing data formats, and aligned with new regulatory expectations is an operational nightmare.

• They’re not reusable. A prompt that works for one document type rarely works for another. This leads to duplicated effort and inconsistent quality.

• They’re not transparent. Prompt chains often obscure where information came from or how it was transformed — a critical issue for regulated content and GxP validation.

And most importantly: prompt libraries don’t scale. Adding a new document type means starting from scratch — new prompts, new edge cases, new integrations. It’s a recipe for slow innovation and user frustration.

The Agentic Advantage

Instead of building ever-larger prompt libraries, Artos built a platform around agents: discrete, composable systems that each perform a specific task in the document authoring workflow. Agents are structured around the core steps of writing a document — data extraction, information retrieval, content reuse, table analysis, summarization, conclusion drafting, QC, and even lean medical writing principles — all in a modular, reusable way.

This agentic structure offers key advantages:

• Modularity. Agents are designed to be reused across document types. Once you build a ‘QC’ agent, it can be used as-is in CSRs, Module 2 Written Summaries, CMC documents, and more.

• Transparency. Users can see exactly what data was used, how it was processed, and why certain content ended up in a draft. AI can show its work.

• Scalability. New document types don’t require net-new engineering. You simply orchestrate existing agents in new ways, guided by example documents and templates.

• Maintainability. If a new LLM model improves summarization, you only need to update the summarization agent once — and every workflow that uses it benefits automatically.

• No prompt engineering required. Users don’t have to guess how to phrase an instruction or understand the quirks of each model. Agents reliably and reproducibly interface with LLMs to produce consistent outputs. This frees domain experts to focus on their actual work, not the internals of AI systems.

Designed for the Real World

Life sciences document workflows are messy. Inputs come in all sorts of formats — .pdf, .rtf, .docx, .jpeg, .xlsx, and more. Content varies by vendor, study, and therapeutic area. Prompt libraries can’t keep up — but agents can.

A robust agentic system for ingestion can transform raw, heterogeneous content into structured, AI-intelligible formats. It works with handwritten notes, scanned reports, and internal company databases (e.g., CMC or clinical data). It knows how to store information across relational databases, vector stores, or both, and how to capture metadata as relevant context — without requiring manual mapping by the user.

A well-designed retrieval agent can then filter on that metadata, reason like a human to extract the right information, and pass it downstream to agents for summarization, conclusion drafting, or formatting — all while maintaining audit trails and high consistency.

Building Agent Native workflows that scale 

Enterprises don’t need more clever prompts. They need systems that evolve with their needs — systems that are transparent, composable, and maintainable at scale. Most of all, they need systems that respect their users’ time.

That’s why Artos chose to go agent-native from day one. We don’t just support enterprise AI workflows. We make them sustainable — and user-friendly.